def compute_rfa()

in data.py [0:0]

• 34 lines of code
• 2 McCabe index (conditional complexity)

def compute_rfa(features, mode='features', k_neighbours=15, distfn='sym', 
	connected=False, sigma=1.0, distlocal='minkowski'):
	"""
	Computes the target RFA similarity matrix. The RFA matrix of
	similarities relates to the commute time between pairs of nodes, and it is
	built on top of the Laplacian of a single connected component k-nearest
	neighbour graph of the data.
	"""
	start = timeit.default_timer()
	if mode == 'features':
		KNN = kneighbors_graph(features,
							   k_neighbours,
							   mode='distance',
							   metric=distlocal,
							   include_self=False).toarray()

		if 'sym' in distfn.lower():
			KNN = np.maximum(KNN, KNN.T)
		else:
			KNN = np.minimum(KNN, KNN.T)    

		n_components, labels = csgraph.connected_components(KNN)

		if connected and (n_components > 1):
			from sklearn.metrics import pairwise_distances
			distances = pairwise_distances(features, metric=distlocal)
			KNN = connect_knn(KNN, distances, n_components, labels)
	else:
		KNN = features    

	if distlocal == 'minkowski':
		# sigma = np.mean(features)
		S = np.exp(-KNN / (sigma*features.size(1)))
		# sigma_std = (np.max(np.array(KNN[KNN > 0])))**2
		# print(sigma_std)
		# S = np.exp(-KNN / (2*sigma*sigma_std))
	else:
		S = np.exp(-KNN / sigma)

	S[KNN == 0] = 0
	print("Computing laplacian...")    
	L = csgraph.laplacian(S, normed=False)
	print(f"Laplacian computed in {(timeit.default_timer() - start):.2f} sec")

	print("Computing RFA...")
	start = timeit.default_timer()
	RFA = np.linalg.inv(L + np.eye(L.shape[0]))
	RFA[RFA==np.nan] = 0.0
	
	print(f"RFA computed in {(timeit.default_timer() - start):.2f} sec")

	return torch.Tensor(RFA)